Improving Human Detection by Long-Term Observation

In this paper we propose a novel human detection method which is based on the existing learning-based method but designed so as to obtain the scene-specific knowledge and utilize it for improving the detection performance. The scene-specific knowledge contains two kinds of information. One of them is additional positive and negative samples that could not be detected by the initial detection method but extracted afterwards by tracking the initial detection results. The other is camera calibration using the size and direction of the detected people in the scene. By this calibration, we can drastically reducing the possibility to incidentally find a pattern which is not a human but looks similar to a human. Experimental results show the effectiveness of the proposed method.